If a stranger cannot repeat your reading from the sentence, you have not named the instrument

Write the app names, the count behind each one and the date, and your finding becomes something a stranger can repeat. Describe the same work with a category word and it becomes a claim about no particular instrument, with no denominator and nothing to point a second reading at. Of the 102 conversations in which we put that question to ChatGPT and to Gemini on 15 October 2026, 98 asked for the apps to be listed by name rather than gathered under the word AI. That is the second most requested item in the whole specification and it is the cheapest one to supply.

The test is one sentence long. Hand your result to somebody with no access to your files and ask whether they could take the same reading tomorrow. If the answer needs a conversation, the naming is the part that is missing. We sell this measurement, so our own sentences carry the surfaces and the counts in the same breath.

What the category word removes

Three things disappear at once, and each of them is load bearing.

• The count. A reading of ChatGPT and a reading of Gemini have separate totals. Rolled together under one word there is no denominator for either, so the figure cannot be checked against anything, including its own next version.
• The structure. The surfaces do not source alike, so a merged description is a description of neither.
• The re-run. Somebody repeating your work has to know which products to open. That is the entire difference between a published finding and a published assertion.

The first of the three has arithmetic behind it, which is the quickest way to see what the category word costs. Our run of 15 October 2026 produced 597 usable readings, and that single total is the version most reports would print. Split by product it reads 306 on ChatGPT and 291 on Gemini. The split is where the information is: inside those two columns the most used source on one reached 176 of its own readings and on the other 19 of its own. A reader handed 597 cannot recover either of those fractions. A reader handed 306 and 291 can check both before breakfast.

Naming two apps is not enough on its own, because pairings differ

Even a named pair is specific to itself. This site publishes four separate readings and they do not all use the same two products. The pair behind both citation runs and the city scan was ChatGPT with Gemini. One further reading, a blind scan dated 2026-10-16, used Perplexity in place of Gemini. The overlap between those two pairings is a single product, so a figure produced by one of them cannot be carried into a sentence about the other, and writing the pair out is what makes that obvious instead of leaving it for a reader to work out.

So the naming requirement goes further than listing products. It names the pairing that produced the specific figure in the specific sentence. A register of those four readings, with the question each one may answer, is published as the register of the four readings this site holds.

The sentence that passes, in four parts

A reproducible engine sentence contains four things and none of them is optional.

• The apps, by their product names. The things a person opens, written the way the product writes them.
• A count per app. Not a run total. 306 and 291 tell a reader something that 597 does not.
• The date. Retrieval behaviour moves, so a figure describes the day it was taken.
• The surface, labelled. Whether the reading came from the consumer app or from an interface, because those are two layers of one product and they can answer differently.

Our own strongest result is written to that pattern and it is a zero: across two runs, Gemini produced 921 readings with our domain in none of them, 291 of those on 15 October 2026 and 630 on 16 October 2026. Every part of that sentence is checkable and none of it would survive being compressed into a category word.

Two is not five, in every sentence

The arithmetic that gets fudged most often in this field is the gap between what a product covers and what a dataset covers. Ours reads five surfaces for a customer: ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Each of our citation runs reached two of those five. Those are separate facts and they belong in separate sentences, permanently, including on a page that is about this exact problem.

The honest form is dull and short: name the dataset's surfaces where you give the dataset's figures, and name the product's coverage where you describe the product. A reader who finds the two merged has found the most common overclaim in the category, and it is usually accidental rather than deliberate, which is why a house rule is worth more than care.

The same discipline applies to documents, not only to apps

A restated document goes stale into a false claim in exactly the way an unnamed product list goes vague into an unrepeatable one. On 16 October 2026 we read an operator's own crawler documentation and found four identities listed where two pages of ours had said three, with one of those pages having turned the count into an assertion. Both were repaired, and the repair included writing down when the list had been read.

The repair is the same in both cases: name the thing, print the date you read it, and put a re-read in the diary. Which documents settle which questions, named individually rather than referred to as guidance, is worked through in the documentation Google publishes itself and what each document settles.

What this count is and is not

The figure records how many of 102 saved conversations raised the same request, on one day, with two products in the room. We did not test whether naming the apps raises a citation rate, and this count does not measure it. It is not a ranking factor. What it does establish is priority: inside a specification of ten items, the engines put the product names above the sample size and above the published question set.

The conversations themselves cover two surfaces. Perplexity, Claude and Google AI Overviews were not asked, so the count describes what two products wanted rather than a consensus across the field.

Questions about naming the apps in a visibility finding

Is it really worth the extra words?

The names are shorter than the category phrase in most sentences, and they turn an assertion into something repeatable. 98 of 102 conversations asked for them, which makes this the cheapest item in the specification by a wide margin.

What if a client only wants one headline number?

Give the headline with its surfaces named in the same line, then the per surface rows underneath. A single figure across two products hides whichever column is the zero, and the zero column is the one that tells anybody where next quarter's work goes.

How do I name a surface that has several modes?

Name the product and then the mode you used, and keep it identical between readings. A search enabled answer and a recalled answer are different products of the same application, and changing mode between rounds produces a movement somebody will read as a result.

Can I compare my named two surface reading with somebody else's?

Only where the surfaces, the counting rule, the question set and the dates line up, which is rare. The comparison that always works is your own second reading against your own first, on frozen wording, which is why the naming matters most for your own records.

What should a report say about the surfaces it did not read?

Name them as absent. A list of three products that were not asked is more informative than silence, and it stops a reader assuming a coverage the reading never had.

The names are what make a result a measurement

The rest of a visibility figure is arithmetic anybody can check: a numerator, a denominator, a counting rule, a date. The names are the part that tells a sceptic where to go and do the checking, and without them the arithmetic has nothing to be about. Naming costs nothing, which is the awkward part of this whole subject. The reading itself is the thing we charge for, and the figures we publish about our own domain are on aiknowsus.com/proof, in the company of two exports that were not chosen for how they look.